v0.1.0
Changes:
- Six new data sources! (see below)
- Documentation is revamped to make it easier for new users to get started
- Fix pip dependency resolution issues so users can install rslearn when using pip directly
- Support multi-temporal Items and add SPATIAL_MOSAIC_TEMPORAL_STACK compositing method to handle aligning item time series against the window time range
- Improved behavior for the dataset prepare/ingest/materialize commands: use multiple workers by default, and process all windows (instead of failing on errors) while showing a summary of processed windows and errors at the end
- RegressionTask and PerPixelRegressionTask: add RMSE and MAPE metrics
- OlmoEarth: support training with missing modalities, and expose sub-month missing timesteps
- OlmoEarth: use olmoearth_pretrain_minimal to handle ModelIDs so users can easily install it via rslearn[extra]
- New EmbeddingCache model component: caches embeddings when using a frozen encoder so they don't need to be re-computed each time; makes it much faster to train a linear probe
- Support easily moving windows between groups via renaming (previously, this would cause issues since the group was also stored in metadata.json)
- During data loading, make windows skipped by check_window more transparent
- New SQLiteWindowStorage window storage: much faster window listing and completed layer lookup operations than the default FileWindowStorage
- New transforms: GaussianNoise, RandomTimeDropping
- Lots of bug fixes and test coverage improvements
New data sources:
- SoilDB
- Sentinel-2 from EarthDaily
- ERA5 Land Hourly Time-Series
- Sentinel-3 SLSTR LST from Planetary Computer
- Google Satellite Embedding v1 from AWS S3
- HLS Sentinel-2 and Landsat data from Planetary Computer